metadata
license: cc-by-nc-4.0
language:
- en
- he
base_model: Hebrish/Sabich
base_model_relation: finetune
pipeline_tag: automatic-speech-recognition
library_name: nemo
tags:
- parakeet
- tdt
- english
- hebrew
- code-switching
- dictation
- continued-training
extra_gated_prompt: >-
Access to this model is by request and is governed by the Hebrish terms of use
(https://hebrish.app/terms). Non-commercial use is free; commercial use needs
a commercial licence from access@hebrish.app.
extra_gated_fields:
Name: text
Organisation: text
Intended use: text
I agree to use this model under the Hebrish terms of use (non-commercial unless I hold a commercial licence): checkbox
Sabich (Rolling) r1
Sabich (Rolling) is Sabich trained forward: each release continues
from the previous weights on all the recordings collected so far, instead of retraining from stock
nvidia/parakeet-tdt-0.6b-v2. It is the default model in the Hebrish apps from
2026-11-02 while it is tested; Sabich stays available.
English dictation with Hebrew (and some Yiddish) words dropped in, written in Latin letters, plus tech vocabulary, numbers and Israeli names. Same tokenizer, output style and usage as Sabich.
r1 (2026-11-02)
- Initialised from Sabich v2.4, continued for 6 epochs at learning rate 1e-5 on the full Sabich training blend (about 26 hours of audio), fresh optimizer, 10% public-English replay, last-5 checkpoint averaging.
- Checked internally against Sabich v2.4 on the same held-out sets before release.
Files
Sabich-Rolling.nemo: the NeMo model (ASRModel.restore_from("Sabich-Rolling.nemo")).Hebrish/Sabich-Rolling-sherpa-onnx-int8: int8 sherpa-onnx build used by the apps.
Contact: access@hebrish.app